Documentation

Ponderly turns thinking out loud with an AI into a map you can walk afterwards. Your own AI does the writing, over MCP, while you talk. This page covers connecting one, what it can do, and what happens to everything it writes.

What Ponderly is

While you explore an idea with Claude or ChatGPT, an AI grows a knowledge tree alongside the conversation. Subjects become nodes. Containment becomes branches. The jumps between ideas become tangents. It happens live, as the conversation wanders, and it costs you nothing to say.

Afterwards the tree is a navigable record of what you explored — searchable, zoomable, and carrying the provenance of when and where each thought happened.

The intelligence is your AI, not a feature we host. Ponderly runs no language model. It is the store and the map; your LLM does the thinking and writes to the tree as you talk.

Connect your AI

Ponderly is a remote MCP server. Add it once in whichever app you talk to, and everything below becomes available to that AI.

MCP server URL
https://api.ponderly.ai/mcp

Claude

  1. Open Customize → Connectors.
  2. Click +, then Add custom connector. Name it Ponderly and paste the address above as the remote MCP server URL.
  3. Authorize with your Ponderly account when prompted.
  4. Open the connector and set Tool permissions to Always allow — both groups. This is the step that decides whether Ponderly is invisible or in your way.
  5. In any new conversation, say "set up Ponderly". Claude installs what it needs and confirms.
Both groups, set once. The two counts come from the tools themselves — Ponderly tells Claude which of them only read.

Do it now rather than when asked. Claude's approval dialog puts Allow once in the highlighted button, so the quickest answer is the one that makes it ask again on the next call — and Ponderly writes every time you follow a thought somewhere.

One connector covers every Claude surface: claude.ai, Desktop, mobile and Claude Code all read the same catalog. Add it on claude.ai or Desktop, though. The iOS and Android apps can use a connector that is already on your account, but adding one there is still in beta.

ChatGPT

  1. Turn on developer mode first, under Settings → Plugins → Advanced. Custom MCP connectors do not appear until it is on — this is the step people miss.
  2. Click Create, name it Ponderly, and use the address above as the MCP server URL. Keep the /mcp path.
  3. In any new chat, say "set up Ponderly".

ChatGPT has no installable skill files on Free, Plus or Pro, so paste the exploring-ideas skill that setup returns into Settings → Personalization → Custom instructions. Business, Enterprise, Healthcare and Edu plans do have skill uploads, and on Enterprise and Edu an admin may need to turn Skills on first. Either way, the skills page has every skill ready to paste or to download. ChatGPT also has no standing pre-approval: it asks before each write, and choosing to remember your answer holds for that conversation only. A new chat starts asking again. That is ChatGPT rather than Ponderly, and Claude is the quieter home for this today.

You need an account first

Sign up at ponderly.ai with Google or a phone number. It is free, and there is no configuration — no API key to generate, no workspace to provision.

How writing works

There is no handshake before a write. Your AI asks in the moment, in its own words, whether to log something, and the write either lands or it does not.

  • Reading is free. An AI can look at your forest any time, without asking, which is what lets it pick up a thread from three weeks ago without making a ceremony of it.
  • Writing is not gated on our side. Your AI is expected to ask before it logs something, and its Ponderly instructions tell it to. That expectation lives in the instructions, not in a check the server makes: anything you have connected can write to your forest.
  • It is never homework. Writes land live. There is no queue of suggestions to approve, no pending state, nothing waiting for you later.
  • Nothing is destroyed. Every removal or replacement keeps what was there before, and can be undone.

If you would rather it stopped, say so and it will drop the subject for the rest of the conversation. To stop it for good, remove the Ponderly connector from that AI. There is no switch here that holds writes back once something is connected.

The words

These are used consistently everywhere — in the app, in the tools, and in anything your AI says back to you.

Branches say what belongs under what. Tangents say what led to what, and they are free to cross between trees.
node
A subject. The unit everything else hangs off.
branch
The parent edge — "belongs under". Every node has at most one, so the skeleton stays a strict tree.
tangent
A free cross-link carrying the thought that connected two nodes — "led to". Unconstrained: any two nodes, any trees, any depth.
root / tree / forest
A node with no parent, everything under it, and all of them together.
visit
One entry in a node's history: when you were there, and what you were doing.
reference
A link out to a book, doc, video or page, attached to a node.
sketch
A visual note on a node — a self-contained HTML document your AI writes and the app renders sandboxed.

What your AI can do

Fourteen tools. Five only look; the other nine write, and none of them destroys anything.

Readingnever asks

read_forest
Every tree at a glance, with explicit markers where it had to truncate. Name a subject and it also points at the regions that already cover it.
read_tree
One branch and everything under it, plus the path back to its root.
read_idea
One idea in full: its description, its links out, everything it connects to, and every time you came back to it.
search_forest
Semantic and keyword search together, over your own forest and nothing else.
get_skills
How an AI keeps its own instructions current: which versions exist, and the full text of any one of them.

Capturingwrites

create_idea
Makes a new idea and says where it goes. Placement is always explicit; the server never guesses.
visit_idea
Records that you came back to something, and what brought you there. Nothing new is made.
add_reference
Hangs a book, article, video or link on an idea.
set_sketch
Sets or replaces an idea's visual note.

Reshapingwrites

edit_idea
Renames it, adds to its description, rewrites it, or moves it somewhere else. The previous text is kept.
create_tangent
Connects two ideas that are not parent and child, carrying the thought that connected them.
merge_ideas
Folds one idea into another. Your AI only does this after you have said so.
prune
Takes an idea, a connection or a reference off the map. Nothing is hard-deleted.

Setupwrites

setup
Onboarding in one call: the skills, plus install instructions for whichever AI you are using.

You do not call these. Your AI does, in the course of talking to you — the names are here so that when it tells you what it did, the sentence means something.

Your forest on the web

Everything your AI writes shows up at ponderly.ai/me, as a map you can zoom and pan. Pull the camera back and nodes decay to bare dots; push in and each one blooms into a card you can read.

  • Tend it by hand. Rename, re-parent and prune directly. Your AI proposes; you are the authority, always, and you never have to ask it to fix something.
  • Open a node for its body, its references, its sketch, and every visit it has ever had.

Sharing, and privacy of it

Your profile lives at ponderly.ai/<your handle>, and you choose whether anyone can walk it.

  • Public — anyone with the address can see your profile and explore your ideas.
  • Private — only people you approve. Your name and handle stay visible so people can ask.

You can change your mind whenever, in Settings. Nodes and sketches behind a private profile answer to a stranger exactly as if they had never existed.

What we store, and what we never see

The short version: we hold the map your AI draws, and almost nothing else.

We never receive your conversations

Ponderly makes no model calls. The only model on our servers is a small embedding model used as a search index, and it never judges anything. What reaches us is what your AI explicitly writes — a node name, a body, the context of a jump — not the transcript around it. Nothing you keep here is used to train anything.

What is stored

Your account record (sign-in identity, name, handle, email or phone), nodes and their body history, edges, visits, references, sketches, images you upload, your follows and invites, and a log of which tools were called.

Scope

A connected AI's access is scoped to one account: your own forest, and nothing else. There is no path from one account to another.

Who else touches it

Firebase for sign-in, Neon for the database, Cloudflare R2 for images, Google Cloud Run for the API, Vercel for the web app. No one else, and no advertising or analytics broker.

The full detail is in the privacy policy, and the terms say what we each owe the other.

Getting help

Write to hello@bycrux.com — a person reads it. Tell us what you were doing and what happened instead; if it involves a specific node, the address of the page is the fastest thing you can send.